handle execveat systemcall on McKernel

Refs: #1366
Change-Id: I921e04a0df8d0d798fc94f675e5112dd2fec190a

arch/arm64/kernel/include/syscall_list.h

@@ -119,6 +119,7 @@ SYSCALL_DELEGATED(241, perf_event_open)
 SYSCALL_HANDLED(260, wait4)
 SYSCALL_HANDLED(270, process_vm_readv)
 SYSCALL_HANDLED(271, process_vm_writev)
+SYSCALL_HANDLED(281, execveat)
 SYSCALL_HANDLED(700, get_cpu_id)
 #ifdef PROFILE_ENABLE
 SYSCALL_HANDLED(__NR_profile, profile)

arch/x86_64/kernel/include/syscall_list.h

@@ -163,6 +163,7 @@ SYSCALL_HANDLED(303, mod_call)
 SYSCALL_HANDLED(309, getcpu)
 SYSCALL_HANDLED(310, process_vm_readv)
 SYSCALL_HANDLED(311, process_vm_writev)
+SYSCALL_HANDLED(322, execveat)
 SYSCALL_HANDLED(700, get_cpu_id)
 #ifdef PROFILE_ENABLE
 SYSCALL_HANDLED(__NR_profile, profile)

executer/user/mcexec.c

@@ -3521,6 +3521,39 @@ again:
 	return ret;
 }
 
+/* for execveat */
+static int getpath_execveat(int dirfd, const char *filename, int flags,
+		char *pathbuf, size_t size)
+{
+	int rc, ret = 0;
+	size_t len;
+
+	if (filename[0] == '/' || dirfd == AT_FDCWD) {
+		len = snprintf(pathbuf, size, "%s", filename);
+	}
+	else if (flags & AT_EMPTY_PATH && filename[0] == '\0') {
+		len = snprintf(pathbuf, size, "/dev/fd/%d", dirfd);
+	}
+	else {
+		len = snprintf(pathbuf, size, "/dev/fd/%d/%s", dirfd, filename);
+	}
+
+	if (len >= size) {
+		ret = ENAMETOOLONG;
+		goto out;
+	}
+
+	if (flags & AT_SYMLINK_NOFOLLOW) {
+		if ((rc = readlink(filename, pathbuf, PATH_MAX)) != -1) {
+			ret = ELOOP;
+			goto out;
+		}
+	}
+
+out:
+	return ret;
+}
+
 int main_loop(struct thread_data_s *my_thread)
 {
 	struct syscall_wait_desc w;
@@ -3929,6 +3962,7 @@ fork_err:
 			break;
 		}
 
+		/* Actually, performing execveat() for McKernel */
 		case __NR_execve: {
 
 			/* Execve phase */
@@ -3940,15 +3974,25 @@ fork_err:
 				char *shebang_argv_flat;
 				char *buffer;
 				size_t size;
-				int ret;
+				int ret, dirfd, flags;
 
 				/* Load descriptor phase */
 				case 1:
 					shebang_argv = NULL;
 					buffer = NULL;
 					desc = NULL;
-					filename = (char *)w.sr.args[1];
+					dirfd = (int)w.sr.args[1];
+					filename = (char *)w.sr.args[2];
+					flags = (int)w.sr.args[4];
 					
+					ret = getpath_execveat(dirfd,
+						filename, flags,
+						pathbuf, PATH_MAX);
+					if (ret) {
+						goto return_execve1;
+					}
+					filename = pathbuf;
+
 					if ((ret = load_elf_desc_shebang(filename, &desc,
 									 &shebang_argv, 0)) != 0) {
 						goto return_execve1;
@@ -3987,7 +4031,7 @@ fork_err:
 
 					/* Copy descriptor to co-kernel side */
 					trans.userp = buffer;
-					trans.rphys = w.sr.args[2];
+					trans.rphys = w.sr.args[3];
 					trans.size = size;
 					trans.direction = MCEXEC_UP_TRANSFER_TO_REMOTE;
 					

kernel/include/syscall.h

@@ -122,6 +122,11 @@
 # define CLONE_NEWNET	0x40000000	/* New network namespace.  */
 # define CLONE_IO	0x80000000	/* Clone I/O context.  */
 
+/* for execveat() */
+#define AT_FDCWD	-100
+#define AT_SYMLINK_NOFOLLOW	0x100
+#define AT_EMPTY_PATH	0x1000
+
 struct user_desc {
 	unsigned int  entry_number;
 	unsigned int  base_addr;

kernel/syscall.c

@@ -2493,13 +2493,11 @@ static void munmap_all(void)
 	return;
 } /* munmap_all() */
 
-SYSCALL_DECLARE(execve)
+static int do_execveat(ihk_mc_user_context_t *ctx, int dirfd,
+		const char *filename, char **argv, char **envp, int flags)
 {
 	int error;
 	long ret;
-	const char *filename = (const char *)ihk_mc_syscall_arg0(ctx);
-	char **argv = (char **)ihk_mc_syscall_arg1(ctx);
-	char **envp = (char **)ihk_mc_syscall_arg2(ctx);
 
 	char *argv_flat = NULL;
 	int argv_flat_len = 0;
@@ -2538,8 +2536,10 @@ SYSCALL_DECLARE(execve)
 	/* Request host to open executable and load ELF section descriptions */
 	request.number = __NR_execve;  
 	request.args[0] = 1;  /* 1st phase - get ELF desc */
-	request.args[1] = (unsigned long)filename;	
+	request.args[1] = dirfd;
-	request.args[2] = virt_to_phys(desc);
+	request.args[2] = (unsigned long)filename;
+	request.args[3] = virt_to_phys(desc);
+	request.args[4] = flags;
 	ret = do_syscall(&request, ihk_mc_get_processor_id());
 
 	if (ret != 0) {
@@ -2626,7 +2626,7 @@ SYSCALL_DECLARE(execve)
 	}		
 
 	/* Request host to transfer ELF image */
-	request.number = __NR_execve;  
+	request.number = __NR_execve;
 	request.args[0] = 2;  /* 2nd phase - transfer ELF image */
 	request.args[1] = virt_to_phys(desc);
 	request.args[2] = sizeof(struct program_load_desc) + 
@@ -2687,6 +2687,14 @@ end:
 	return ret;
 }
 
+SYSCALL_DECLARE(execve)
+{
+	return do_execveat(ctx, AT_FDCWD,
+			(const char *)ihk_mc_syscall_arg0(ctx),
+			(char **)ihk_mc_syscall_arg1(ctx),
+			(char **)ihk_mc_syscall_arg2(ctx), 0);
+}
+
 unsigned long do_fork(int clone_flags, unsigned long newsp,
                       unsigned long parent_tidptr, unsigned long child_tidptr,
                       unsigned long tlsblock_base, unsigned long curpc,
@@ -3569,6 +3577,39 @@ out:
 	return rc;
 }
 
+SYSCALL_DECLARE(execveat)
+{
+	int dirfd = (int)ihk_mc_syscall_arg0(ctx);
+	const char *filename = (const char *)ihk_mc_syscall_arg1(ctx);
+	int flags = (int)ihk_mc_syscall_arg4(ctx);
+	long ret;
+
+	/* validate flags */
+	if ((flags & ~(AT_SYMLINK_NOFOLLOW | AT_EMPTY_PATH)) != 0) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	if (filename[0] == '/' || dirfd == AT_FDCWD) {
+		/* behave same as execve */
+		goto exec;
+	}
+
+	/* validate dirfd */
+	if (dirfd < 0 && dirfd != AT_FDCWD) {
+		ret = -EBADF;
+		goto out;
+	}
+
+exec:
+	ret = do_execveat(ctx, dirfd, filename,
+			(char **)ihk_mc_syscall_arg2(ctx),
+			(char **)ihk_mc_syscall_arg3(ctx), flags);
+
+out:
+	return ret;
+}
+
 SYSCALL_DECLARE(close)
 {
 	int fd = ihk_mc_syscall_arg0(ctx);

test/issues/1366/C1366.sh

@@ -0,0 +1,29 @@
+#/bin/sh
+
+USELTP=1
+USEOSTEST=0
+
+. ../../common.sh
+
+issue=1366
+tid=01
+
+cp ${LTPBIN}/execve_* ./
+cp ${LTPBIN}/execveat_* ./
+
+for tp in execveat01 execveat02 execveat03 execve01 execve02 execve03 execve05
+do
+	tname=`printf "C${issue}T%02d" ${tid}`
+	echo "*** ${tname} start *******************************"
+	sudo PATH=${LTPBIN}:${PATH} $MCEXEC $LTPBIN/$tp 2>&1 | tee $tp.txt
+	ok=`grep PASS $tp.txt | wc -l`
+	ng=`grep FAIL $tp.txt | wc -l`
+	if [ $ng = 0 ]; then
+		echo "*** ${tname} PASSED ($ok)"
+	else
+		echo "*** ${tname} FAILED (ok=$ok ng=%ng)"
+	fi
+	let tid++
+	echo ""
+done
+

test/issues/1366/Makefile

@@ -0,0 +1,11 @@
+CFLAGS=-g
+LDFLAGS=
+
+TARGET=
+
+all: $(TARGET)
+
+test: all
+	./C1366.sh
+clean:
+	rm -f $(TARGET) *.o *.txt execve_* execveat_*

test/issues/1366/README

@@ -0,0 +1,24 @@
+【Issue#1366 動作確認】
+□ テスト内容
+1. 以下のLTPを用いてexecveat機能がMckernelで動作することを確認
+	- execveat01
+	- execveat02
+	- execveat03
+
+2. 以下のLTPを用いて既存のexecve機能に影響が無いことを確認
+	- execve01
+	- execve02
+	- execve03
+	- execve05
+
+□ 実行手順
+$ make test
+
+McKernelのインストール先や、OSTEST, LTPの配置場所は、
+$HOME/.mck_test_config を参照している
+.mck_test_config は、McKernelをビルドした際に生成されるmck_test_config.sample ファイルを
+$HOMEにコピーし、適宜編集する
+
+□ 実行結果
+x86_64_result.log aarch64_result.log 参照。
+すべての項目をPASSしていることを確認。

test/issues/1366/aarch64_result.log

@@ -0,0 +1,95 @@
+*** C1366T01 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execveat_child.c:36: PASS: execveat_child run as expected
+execveat_child.c:36: PASS: execveat_child run as expected
+execveat_child.c:36: PASS: execveat_child run as expected
+execveat_child.c:36: PASS: execveat_child run as expected
+
+Summary:
+passed   4
+failed   0
+skipped  0
+warnings 0
+*** C1366T01 PASSED (4)
+
+*** C1366T02 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execveat02.c:64: PASS: execveat() fails as expected: EBADF
+execveat02.c:64: PASS: execveat() fails as expected: EINVAL
+execveat02.c:64: PASS: execveat() fails as expected: ELOOP
+execveat02.c:64: PASS: execveat() fails as expected: ENOTDIR
+
+Summary:
+passed   4
+failed   0
+skipped  0
+warnings 0
+*** C1366T02 PASSED (4)
+
+*** C1366T03 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execveat_child.c:36: PASS: execveat_child run as expected
+
+Summary:
+passed   1
+failed   0
+skipped  0
+warnings 0
+*** C1366T03 PASSED (1)
+
+*** C1366T04 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execve01_child.c:46: PASS: execve01_child executed
+
+Summary:
+passed   1
+failed   0
+skipped  0
+warnings 0
+*** C1366T04 PASSED (1)
+
+*** C1366T05 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execve02.c:66: PASS: execve() failed expectedly: EACCES
+
+Summary:
+passed   1
+failed   0
+skipped  0
+warnings 0
+*** C1366T05 PASSED (1)
+
+*** C1366T06 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execve03.c:144: PASS: execve failed as expected: ENAMETOOLONG
+execve03.c:144: PASS: execve failed as expected: ENOENT
+execve03.c:144: PASS: execve failed as expected: ENOTDIR
+execve03.c:144: PASS: execve failed as expected: EFAULT
+execve03.c:144: PASS: execve failed as expected: EACCES
+execve03.c:144: PASS: execve failed as expected: ENOEXEC
+
+Summary:
+passed   6
+failed   0
+skipped  0
+warnings 0
+*** C1366T06 PASSED (6)
+
+*** C1366T07 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+
+Summary:
+passed   8
+failed   0
+skipped  0
+warnings 0
+*** C1366T07 PASSED (8)
+

test/issues/1366/x86_64_result.log

@@ -0,0 +1,95 @@
+*** C1366T01 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execveat_child.c:36: PASS: execveat_child run as expected
+execveat_child.c:36: PASS: execveat_child run as expected
+execveat_child.c:36: PASS: execveat_child run as expected
+execveat_child.c:36: PASS: execveat_child run as expected
+
+Summary:
+passed   4
+failed   0
+skipped  0
+warnings 0
+*** C1366T01 PASSED (4)
+
+*** C1366T02 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execveat02.c:64: PASS: execveat() fails as expected: EBADF
+execveat02.c:64: PASS: execveat() fails as expected: EINVAL
+execveat02.c:64: PASS: execveat() fails as expected: ELOOP
+execveat02.c:64: PASS: execveat() fails as expected: ENOTDIR
+
+Summary:
+passed   4
+failed   0
+skipped  0
+warnings 0
+*** C1366T02 PASSED (4)
+
+*** C1366T03 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execveat_child.c:36: PASS: execveat_child run as expected
+
+Summary:
+passed   1
+failed   0
+skipped  0
+warnings 0
+*** C1366T03 PASSED (1)
+
+*** C1366T04 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execve01_child.c:46: PASS: execve01_child executed
+
+Summary:
+passed   1
+failed   0
+skipped  0
+warnings 0
+*** C1366T04 PASSED (1)
+
+*** C1366T05 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execve02.c:66: PASS: execve() failed expectedly: EACCES
+
+Summary:
+passed   1
+failed   0
+skipped  0
+warnings 0
+*** C1366T05 PASSED (1)
+
+*** C1366T06 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execve03.c:144: PASS: execve failed as expected: ENAMETOOLONG
+execve03.c:144: PASS: execve failed as expected: ENOENT
+execve03.c:144: PASS: execve failed as expected: ENOTDIR
+execve03.c:144: PASS: execve failed as expected: EFAULT
+execve03.c:144: PASS: execve failed as expected: EACCES
+execve03.c:144: PASS: execve failed as expected: ENOEXEC
+
+Summary:
+passed   6
+failed   0
+skipped  0
+warnings 0
+*** C1366T06 PASSED (6)
+
+*** C1366T07 start *******************************
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+execve_child.c:34: PASS: argv[1] is canary, expected 'canary'
+
+Summary:
+passed   8
+failed   0
+skipped  0
+warnings 0
+*** C1366T07 PASSED (8)
+